[PATCH] Add a REG_P check for inc and dec for Arm MVE

[PATCH] Add a REG_P check for inc and dec for Arm MVE

[Saurabh Jha]

[-- Attachment #1: Type: text/plain, Size: 590 bytes --]

Hey,

This patch tightens mve_vector_mem_operand to reject non-register operands inside {PRE,POST}_{INC,DEC} addresses by introducing a REG_P check.

This patch fixes this ICE:https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112337

Okay for trunk? I don't have trunk access so could someone please commit on my behalf?

Regards,
Saurabh

gcc/ChangeLog:

	PR target/112337
	* config/arm/arm.cc (mve_vector_mem_operand): Add a REG_P check for INC
	and DEC operations

gcc/testsuite/ChangeLog:

	PR target/112337
	* gcc.target/arm/mve/pr112337.c: Test for REG_P check for INC and DEC
	operations

[-- Attachment #2: rb17987.patch --]
[-- Type: text/plain, Size: 1877 bytes --]

From a53553878602c7050b4f027a19149ca643b11721 Mon Sep 17 00:00:00 2001
From: Saurabh Jha <saujha01@e130340.arm.com>
Date: Wed, 8 Nov 2023 14:33:35 +0000
Subject: [PATCH 1/1] Add a REG_P check for inc and dec for Arm MVE

It is okay to have operands in MEM for inc and dec in Arm
MVE. This patch introduces a REG_P check for inc and dec so that
we don't get an ICE for the case of inc/dec on MEM.
---
 gcc/config/arm/arm.cc                       |  7 +++++--
 gcc/testsuite/gcc.target/arm/mve/pr112337.c | 15 +++++++++++++++
 2 files changed, 20 insertions(+), 2 deletions(-)
 create mode 100644 gcc/testsuite/gcc.target/arm/mve/pr112337.c

diff --git a/gcc/config/arm/arm.cc b/gcc/config/arm/arm.cc
index 620ef7bfb2f..25a1ad736ad 100644
--- a/gcc/config/arm/arm.cc
+++ b/gcc/config/arm/arm.cc
@@ -13695,8 +13695,11 @@ mve_vector_mem_operand (machine_mode mode, rtx op, bool strict)
     }
   code = GET_CODE (op);
 
-  if (code == POST_INC || code == PRE_DEC
-      || code == PRE_INC || code == POST_DEC)
+  if ((code == POST_INC
+       || code == PRE_DEC
+       || code == PRE_INC
+       || code == POST_DEC)
+      && REG_P (XEXP (op, 0)))
     {
       reg_no = arm_effective_regno (XEXP (op, 0), strict);
       return (((mode == E_V8QImode || mode == E_V4QImode || mode == E_V4HImode)
diff --git a/gcc/testsuite/gcc.target/arm/mve/pr112337.c b/gcc/testsuite/gcc.target/arm/mve/pr112337.c
new file mode 100644
index 00000000000..8f491990088
--- /dev/null
+++ b/gcc/testsuite/gcc.target/arm/mve/pr112337.c
@@ -0,0 +1,15 @@
+/* { dg-do compile } */
+/* { dg-options "-O2" } */
+/* { dg-require-effective-target arm_v8_1m_mve_ok } */
+/* { dg-add-options arm_v8_1m_mve } */
+#include <arm_mve.h>
+
+void g(int32x4_t);
+void f(int, int, int, short, int *p) {
+  int *bias = p;
+  for (;;) {
+    int32x4_t d = vldrwq_s32 (p);
+    bias += 4;
+    g(d);
+  }
+}
-- 
2.34.1

RE: [PATCH] Add a REG_P check for inc and dec for Arm MVE

[Kyrylo Tkachov]

Hi Saurabh,

> -----Original Message-----
> From: Saurabh Jha <saurabh.jha@arm.com>
> Sent: Thursday, November 9, 2023 10:12 AM
> To: gcc-patches@gcc.gnu.org; Richard Earnshaw
> <Richard.Earnshaw@arm.com>; Richard Sandiford
> <Richard.Sandiford@arm.com>
> Subject: [PATCH] Add a REG_P check for inc and dec for Arm MVE
> 
> Hey,
> 
> This patch tightens mve_vector_mem_operand to reject non-register
> operands inside {PRE,POST}_{INC,DEC} addresses by introducing a REG_P
> check.
> 
> This patch fixes this ICE:https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112337
> 
> Okay for trunk? I don't have trunk access so could someone please commit
> on my behalf?

Ok.

> 
> Regards,
> Saurabh
> 
> gcc/ChangeLog:
> 
> 	PR target/112337
> 	* config/arm/arm.cc (mve_vector_mem_operand): Add a REG_P
> check for INC
> 	and DEC operations
> 
> gcc/testsuite/ChangeLog:
> 
> 	PR target/112337
> 	* gcc.target/arm/mve/pr112337.c: Test for REG_P check for INC and
> DEC
> 	operations

ChangeLog entries should end with a full stop (the git commit hooks enforce it).
I've adjusted the ChangeLog and pushed this patch for you.
Thank you for the patch!
Kyrill

Re: [PATCH] Add a REG_P check for inc and dec for Arm MVE

[Christophe Lyon]

Hi Saurabh,

On Tue, 14 Nov 2023 at 15:51, Kyrylo Tkachov <Kyrylo.Tkachov@arm.com> wrote:
>
> Hi Saurabh,
>
> > -----Original Message-----
> > From: Saurabh Jha <saurabh.jha@arm.com>
> > Sent: Thursday, November 9, 2023 10:12 AM
> > To: gcc-patches@gcc.gnu.org; Richard Earnshaw
> > <Richard.Earnshaw@arm.com>; Richard Sandiford
> > <Richard.Sandiford@arm.com>
> > Subject: [PATCH] Add a REG_P check for inc and dec for Arm MVE
> >
> > Hey,
> >
> > This patch tightens mve_vector_mem_operand to reject non-register
> > operands inside {PRE,POST}_{INC,DEC} addresses by introducing a REG_P
> > check.
> >
> > This patch fixes this ICE:https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112337
> >
> > Okay for trunk? I don't have trunk access so could someone please commit
> > on my behalf?
>
> Ok.
>
> >
> > Regards,
> > Saurabh
> >
> > gcc/ChangeLog:
> >
> >       PR target/112337
> >       * config/arm/arm.cc (mve_vector_mem_operand): Add a REG_P
> > check for INC
> >       and DEC operations
> >
> > gcc/testsuite/ChangeLog:
> >
> >       PR target/112337
> >       * gcc.target/arm/mve/pr112337.c: Test for REG_P check for INC and
> > DEC
> >       operations
>

This new test fails in our CI (various flavours of target arm-eabi),
with the following:
/gcc.target/arm/mve/pr112337.c:11:18: warning: passing argument 1 of
'__arm_vldrwq_s32' from incompatible pointer type
[-Wincompatible-pointer-types]
In file included from
/home/tcwg-buildslave/workspace/tcwg_gnu_0/abe/snapshots/gcc.git~master/gcc/testsuite/gcc.target/arm/mve/pr112337.c:5:
/arm-eabi/14.0.0/include/arm_mve.h:1752:35: note: expected 'const
int32_t *' {aka 'const long int *'} but argument is of type 'int *'

I think you should make p a int32_t pointer rather than int, but I am
wondering why you didn't see this problem in your testing?

Can you check?

Thanks,

Christophe

> ChangeLog entries should end with a full stop (the git commit hooks enforce it).
> I've adjusted the ChangeLog and pushed this patch for you.
> Thank you for the patch!
> Kyrill
>

Re: [PATCH] Add a REG_P check for inc and dec for Arm MVE

[Saurabh Jha]

Hi Christophe,

On 11/20/2023 10:12 AM, Christophe Lyon wrote:
> Hi Saurabh,
>
> On Tue, 14 Nov 2023 at 15:51, Kyrylo Tkachov <Kyrylo.Tkachov@arm.com> wrote:
>> Hi Saurabh,
>>
>>> -----Original Message-----
>>> From: Saurabh Jha <saurabh.jha@arm.com>
>>> Sent: Thursday, November 9, 2023 10:12 AM
>>> To: gcc-patches@gcc.gnu.org; Richard Earnshaw
>>> <Richard.Earnshaw@arm.com>; Richard Sandiford
>>> <Richard.Sandiford@arm.com>
>>> Subject: [PATCH] Add a REG_P check for inc and dec for Arm MVE
>>>
>>> Hey,
>>>
>>> This patch tightens mve_vector_mem_operand to reject non-register
>>> operands inside {PRE,POST}_{INC,DEC} addresses by introducing a REG_P
>>> check.
>>>
>>> This patch fixes this ICE:https://gcc.gnu.org/bugzilla/show_bug.cgi?id=112337
>>>
>>> Okay for trunk? I don't have trunk access so could someone please commit
>>> on my behalf?
>> Ok.
>>
>>> Regards,
>>> Saurabh
>>>
>>> gcc/ChangeLog:
>>>
>>>        PR target/112337
>>>        * config/arm/arm.cc (mve_vector_mem_operand): Add a REG_P
>>> check for INC
>>>        and DEC operations
>>>
>>> gcc/testsuite/ChangeLog:
>>>
>>>        PR target/112337
>>>        * gcc.target/arm/mve/pr112337.c: Test for REG_P check for INC and
>>> DEC
>>>        operations
> This new test fails in our CI (various flavours of target arm-eabi),
> with the following:
> /gcc.target/arm/mve/pr112337.c:11:18: warning: passing argument 1 of
> '__arm_vldrwq_s32' from incompatible pointer type
> [-Wincompatible-pointer-types]
> In file included from
> /home/tcwg-buildslave/workspace/tcwg_gnu_0/abe/snapshots/gcc.git~master/gcc/testsuite/gcc.target/arm/mve/pr112337.c:5:
> /arm-eabi/14.0.0/include/arm_mve.h:1752:35: note: expected 'const
> int32_t *' {aka 'const long int *'} but argument is of type 'int *'
>
> I think you should make p a int32_t pointer rather than int, but I am
> wondering why you didn't see this problem in your testing?
>
> Can you check?
>
> Thanks,
>
> Christophe

Thank you for bringing this up. Unfortunately, I misread the test output 
in my local testing. Apologies for that.

I am working on a fix and will be submitting it soon.

Thanks,

Saurabh

>
>> ChangeLog entries should end with a full stop (the git commit hooks enforce it).
>> I've adjusted the ChangeLog and pushed this patch for you.
>> Thank you for the patch!
>> Kyrill
>>